Graduate Student in Artificial Intelligence
Faculty of Engineering
Friedrich-Alexander University Erlangen-Nuremberg
Peterstraße, 90478 Nürnberg, Germany
sujit [dot] debnath [dot] bd [at] gmail [dot] com
+49 157 31158609
[Curriculum Vitae] [Résumé]
Welcome to my portfolio website. I am Sujit Debnath, a graduate student in Artificial Intelligence at Friedrich-Alexander University Erlangen-Nuremberg (FAU), currently living in Nürnberg, Germany. Alongside my academic pursuits, I was also employed as a working student in automation at Elektrobit Automotive GmbH for the past couple of months.
I completed my Bachelor of Science in Computer Science and Engineering (CSE) from North South University (NSU), where I received the distinction magna cum laude. Due to my excellent academic performance, the university awarded me a merit-based tuition waiver from the second semester until my graduation. Additionally, I was among the top 5% of the entire 'Graduating Class of 2019,' which consisted of 3470 graduates. Before my undergraduate studies, I attended the Secondary School Certificate (SSC) examination from Rampura Ekramunnesa Boys' High School and the Higher Secondary Certificate (HSC) examination from Dhaka College.
Prior to my graduate studies, I gained valuable professional experience by working as a Project Coordinator at the Department of Analytics Center of Excellence, Robi Axiata Limited. Additionally, I worked as a Teaching Assistant (TA) and a Lab Instructor at the Department of Electrical and Computer Engineering of NSU immediately after completing my undergraduate studies.
My research interests lie in the field of Artificial Intelligence, particularly in Machine Learning, Natural Language Processing (NLP), and Data Science, particularly finding innovative solutions for complex problems. I firmly believe that data is the driving force behind the digital age, and hence, data science and artificial intelligence have become indispensable in today's world. With my growing passion for ML, data science, and research, I embarked on a research journey in the biomedical field on 'Glaucoma (group of eye conditions) detection' using Fuzzy Logic. Later, during my undergrad ML course, I worked on 'Early Risk Prediction of Stroke,' using various ML algorithms to predict stroke risk. These preliminary research works motivated me to explore research methodologies more closely and pursue further research opportunities. Consequently, my undergraduate thesis on 'Kothamala - A Bangla Speech Recognition Engine' aimed to explore Natural Language Processing (NLP) in Bengali, which is one of the most widely spoken languages globally. We conducted this research-based project in two divisions: Bengali Short Speech Commands Recognition and Bengali Transcription. This research work provided me with a profound understanding of NLP and its implementation in an under-researched language.
Apart from my academic and professional endeavors, I was an active member of the NSU Problem Solvers (NSUPS) community and volunteered for JAAGO Foundation. I also started volunteering at Volunteer for Bangladesh (VBD), Dhaka division in 2014, which is a non-political volunteer and youth wing of JAAGO. Volunteering has taught me valuable lessons, including social responsibility, teamwork, and leadership, which have significantly improved my interpersonal and social skills. However, in my free time, I enjoy reading books of different genres, such as thriller, mystery, fantasy, science fiction, and classic literature.
Through this website, I aim to showcase my academic and professional journey, research work, and extracurricular activities, along with providing a glimpse into my personality and interests. I hope my website reflects my passion for AI and data science and inspires others to explore this exciting field further. I believe that every day is an opportunity to learn something new and improve oneself. Thus, I am constantly evolving myself to take on new challenges and create my own path in life, as famously quoted by George Bernard Shaw,
Life isn't about finding yourself. Life is about creating yourself.
Dissertation Topic: Kothamala - A Bangla Speech Recognition Engine
Shakil Ahmed Sumon, Joydip Chowdhury, Sujit Debnath, Nabeel Mohammed, Sifat Momen
In 1st International Conference on Bangla Speech and Language Processing (ICBSLP), pp. 1-6, 2018
Despite being one of the most extensively spoken languages in the world, the Bengali language has inadequate resources. Due to the lack of adequate and suitable public datasets in Bengali, it is difficult to work with the Bengali Language. In this study, Bengali short speech commands data set has been reported that consists of 10 frequently used short speech commands, each with approximately 100 utterances. Later, three different Convolutional Neural Network (CNN) architectures have been proposed to recognize those short speech commands. One of the models took raw audio as input, whereas another model took Mel-Frequency Cepstral Coefficients (MFCC) of the audio signals as inputs, and the third model leveraged transfer learning by pre-training the model with English short speech commands. Experimental results reveal that the MFCC model shows better accuracy in recognizing Bengali short speech commands where, surprisingly, the model predicting on raw audio data is very competitive. Though the models have shown proficiency in identifying single-syllable words but encounter difficulties in recognizing multi-syllable commands.
Nazmul Alam Diptu, Md. Asif Khan, Sujit Debnath, Abdullah Al Imam, Al Mahadi Hasan Rakib, Kazi Asfaq Ahmed Ador, Rashedur M. Rahman
In 9th International Conference on Intelligent Systems (IS’18), pp. 87-93, 2018
Glaucoma is a group of eye conditions that damage the optic nerve, leading to irreversible vision loss. Hence detecting Glaucoma at an early stage is very crucial for preventing blindness. But, Glaucoma detection often requires an advanced diagnosis that is not accessible to most hospitals in a developing country like Bangladesh. Therefore, the detection of Glaucoma with the help of common and fewer tests would certainly improve Glaucoma patients' condition. In this research, a method is devised using Adaptive Neuro-Fuzzy Inference System (ANFIS) to detect Glaucoma with the data obtained from two ophthalmological tests, which are Optical Coherence Tomography (OCT) and Tonometry along with some other risk factors. Furthermore, the data of two ophthalmological tests are collected from Bashundhara Eye Hospital & Research Institute.
Shakil Ahmed Sumon, Joydip Chowdhury, Sujit Debnath
The Bengali language is one of the most extensively used spoken languages around the world. Nevertheless, the resources in Bengali are inadequate since there is a lack of proper public datasets that can be used in Natural Language Processing (NLP). As a result, no significant NLP research has been conducted in Bengali. Therefore, as the Capstone Project, we conducted a research-based project related to the Bengali language in two divisions; one was Bengali Short Speech Commands Recognition, and the other one was Bengali Transcription. Bengali Short Speech Commands Recognition: In this division, we developed a dataset of popular Bengali short speech commands consisting of 50 frequently used short speech commands. Each command has approximately 200 utterances, which is why it is first of its kind in the context of Bengali datasets. However, each utterance duration is less than 2 seconds, and all the samples were taken in noisy real-life conditions. Later, several experiments had been conducted to recognize ten Bengali short speech commands, and three different Convolutional Neural Network (CNN) architectures had been proposed. One of the models took raw audio as input, whereas another model took Mel-Frequency Cepstral Coefficients (MFCC) of the audio signals as inputs, and the third model leveraged transfer learning by pre-training the model with English short speech commands. Though the model trained with MFCCs performed better, but all the models performed relatively well. Bengali Transcription: The data set used in this division is an open-source dataset taken from OpenSLR, released by Google. It contains 2,18,703 utterances of Bengali speech with their corresponding labels in a CSV file. In this section, combined with a language model, we had developed a Bengali speech to text engine. However, the spectrogram had been incorporated as an audio pre-processing strategy along with MFCCs and raw audios. Later, three different model architectures had been experimented for building speech to text engine. As the audios are sequential inputs, one model had been designed with time distributed dense neural network architecture; another model with Long Short Term Memory (LSTM) architecture combined with convolutional layers to extract features from audios. Additionally, we had trained another minimal model in combination with one convolutional and one time distributed layer. A special kind of loss function, Connectionist Temporal Classification (CTC), which can deal with the variable length of sequences, had been applied as a loss function in all these models. However, given a sequence of audio signals, assuming the audio will be Bengali phonemes; the network will predict a sequence of Bengali characters' probability distribution. However, it was unfortunate that our transcription system was not able to transcript speech into text efficiently.
Sujit Debnath, Md. Asif Khan, Kazi Asfaq Ahmed Ador
Stroke is a life-threatening medical condition that is one of the leading causes of death worldwide. Moreover, prediction and diagnoses of stroke are intricate tasks; require several diagnostic tests costing valuable time and expertise of medical practitioners. A computer-aided system can help doctors minimize the complexity of predicting and diagnosing stroke and take necessary steps by which stroke can be avoided. In this research-based project, several Machine Learning techniques such as Logistic Regression, Support Vector Classifier, etc along with Synthetic Minority Over-sampling Technique (SMOTE) had been used to predict the risk of stroke on 16 different features excluding the class. However, the dataset was initially collected by Farzana et al. from Dhaka Medical College for their research work, which consists of 500 patients' data; among them, 350 patients are diagnosed with a stroke, and 150 are diagnosed as normal or non-stroke. Data were collected from patients' case history, pathological tests, and Lipid profile test. After using the SMOTE to balance the dataset, several classifiers performed notably well, outperforming the other work on the same dataset.